Paying for inference
You pay for tokens, and the invoice does not say which ones. Most of what an application is billed for is text the provider has already seen, sent again because that is how a conversation works, and the price of that re-reading is set by decisions nobody thinks of as cost decisions: how the prompt is assembled, which route the call takes, whether the work could have waited a day.
These pages are for whoever signs off on that bill. They are not about kernels. They give you a number for your own workload, show which line of it moves when you change something, and back it with measurements: 12,795 controlled calls of one model down fourteen serving routes. The why lives in the explainers, when you want it.
- calculatorWhat actually drives my LLM API bill?
Pick a model at list price and a workload shape, then watch the bill split into system prompt, history re-reads, new input and output. Turn caching on and off, and see which line moves.
- calculatorIs it cheaper to self-host?
Published breakevens disagree by more than a hundred times because four assumptions go unstated. Set all four: the workload, the API you are comparing against, the card, and what the busiest hour looks like.
- guideWhat does zero data retention actually cover?
Three different promises get collapsed into one phrase, and the carve-outs are where your data actually is. What the providers' own documents say, and the questions to take into the vendor call.
- guideShould I route everything through OpenRouter?
The same model down 14 serving routes in 12,795 controlled calls: a 6.8x spread in price, one route failing half its calls, cache hit shares from 53% to 100%. What a router buys, what it hides, and where the real discounts are.